A Robust Keypoint-Based Machine Learning Framework for Audio Copy-Move Forgery Detection Using Time-Frequency Representations
IEEE Access, vol.14, pp.141455-141484, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 14
- Publication Date: 2026
- Doi Number: 10.1109/access.2026.3732567
- Journal Name: IEEE Access
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Page Numbers: pp.141455-141484
- Keywords: Audio copy-move forgery, audio forensics, extremely randomized trees, FAST, machine learning, SIFT
- Ataturk University Affiliated: Yes
Abstract
Copy-move forgery is a type of audio forgery performed to conceal or duplicate speech segments. Verifying audio authenticity has become critical in audio forensics. This study proposes a novel machine learning-based copy-move forgery detection framework. Raw audio signals are converted into time-frequency representations and preprocessed. Keypoints are then detected using FAST and described using SIFT. After keypoints are matched and false correspondences eliminated via 1D translation-consensus filtering, a feature set is constructed from the matched keypoints. This feature vector is classified using an Extra Trees classifier to label recordings as forged or authentic. Our approach is evaluated on three datasets consisting of Arabic, Turkish, and English recordings. After determining the most suitable time-frequency representation, experiments are conducted using 5-fold stratified cross-validation for intra-dataset evaluations and cross-dataset testing for inter-dataset evaluations. Under attack-free conditions, stratified 5-fold cross-validation yields F1-scores of 99.25% for DS1 (Arabic), 99.76% for DS2 (Turkish), and 89.04% for DS3 (English). Under attack scenarios, the F1-score does not fall below 96% for DS1, 96.19% for DS2, and 85.83% for DS3. These results show that the method is robust to post-processing attacks and effective across different languages and recording conditions. Comparative results show that this approach outperforms existing methods.